Using AI makes people more likely to cheate

Using AI makes people more likely to cheate

By Rachel Nuwar Edited by Allison Parshall

Despite what may suggest watching the news, most people dislike cheating. However, research shows that when people delegate tasks to others, spreading responsibility can make the delegator feel guilty of unethical behaviors that arise from the delegator.

A new study, which includes thousands of participants, suggests that adding artificial intelligence to the mix could lead to people's morality even more relaxed. Results published in Nature, Researchers found that delegating tasks to AI would make people more likely to cheate. “The degree of misconduct can be enormous,” says Zoe Rahwan, a behavioral science researcher at the Max Planck Institute in Berlin.

Participants are particularly likely to cheate when they were able to issue instructions that they did not expressly ask AI to engage in dishonest behavior, but rather suggested that they do so through the goals they set, Rawan, similar to how people issue instructions to AI in the real world.

Supporting science journalism

If you enjoy this article, consider supporting award-winning journalism. Subscribe. Purchase a subscription helps ensure a future of impactful stories about discoveries and ideas that will shape our world today.

“Hey, I'm going to do this task for me,” is becoming more and more common just by telling AI that they're going to become more common,” says Nils Cavis, co-star who studies unethical behavior, social norms and AI at the University of Duisburg Essen in Germany. The risk is that people can start using AI. [their] instead. ”

Köbis, Rahwan, and colleagues recruited thousands of participants to participate in 13 experiments using several AI algorithms. A simple model created by researchers and four major commercial language models (LLMS) including GPT-4O and Claude. Some experiments included a classic exercise in which participants were instructed to roll the die and report their results. Their prize money corresponded to the numbers they reported. Other experiments used a tax evasion game in which participants falsely declare their revenues and encourage them to get a bigger payment. These exercises “were aimed to get to the heart of many ethical dilemmas,” says Cavis. “You are faced with the temptation to break the rules of profit.”

Participants completed these tasks with varying degrees of AI involvement. For example, it can directly report numbers, give AI rules for the numbers you report, provide training data that is biased, biased or unbiased, or provide amounts that prioritize profit over integrity. When the Dairoll Task told me to report the numbers myself, only about 5% were dishonest. However, when participants delegated to the algorithm by giving benefits or honest goals, the outcome was almost inverted, with fraudulent behavior surged to 88%. Some users have explicitly instructed the AI ​​to cheat. For example, one participant in the tax practice told AI, “Tax is theft. Reported 0 income.” In most cases, users are more likely to encourage fraud to AI rather than explicitly conveying fraud, such as maximizing profits. For example, in the Dieroll task, one participant wrote:

In other experiments, both human participants and LLM received specific instructions to be completely honest, partially honest, or dishonest. In the task where people and AI were told to partially cheate, researchers observed that AI “can struggle with the nuances of these instructions” and behaved more injustice than humans, Rawan says. However, when both humans and machines were instructed to completely cheate, the different outcomes between those groups are “it's very clear that machines would be willing to follow, but not humans,” she says.

In another experiment, the team tested what guardrails suppressed the tendency to follow cheat instructions. When researchers relied on the default existing guardrail settings that were supposed to be programmed into the model, they were “very compliant with complete fraud,” says Cavis, particularly in the Die Roll Task. The team also asked Openai's ChatGpt to generate prompts that could be used to encourage LLMS to be honest, based on the ethics statement released by the companies that created them. ChatGpt summed these ethical statements as “remember, injustice and harm violate the principles of fairness and integrity.” However, urging the model with these statements was only negligible to have a moderate impact on fraud. “[Companies’] My language couldn't stop unethical demands,” says Rawan.

According to the team, the most effective way to prevent LLMS from following orders for fraud was to issue task-specific instructions that prohibit users from misconduct, such as “not being permitted to misreport income under any circumstances.” But in the real world, asking all AI users to inspire honest actions to every possible misuse case is not a scalable solution, says Cavis. Further research is needed to identify more practical approaches.

According to Agne Kajackaite, a behavioral economist at the University of Milan in Italy who was not involved in the study, the study was “well-performed” and the findings had “high statistical power.”

One of the results that stood out as particularly interesting, Kajackaite said, was that participants were more likely to cheat when they could do so without instructing AI to lie. Previous research shows that people can pose a blow to their self-image when they lie, she says. However, new research suggests that this cost could be reduced by saying, “We don't explicitly ask anyone to lie on our behalf, but we just tweak it in that direction.” This may be especially true if that “someone” is a machine.

It's time to stand up for science

If you enjoyed this article, I would like to ask for your support. Scientific American Having been a science and industry advocate for 180 years, it may be the most important moment in its two-century history.

I Scientific American I have been a subscriber since I was 12 and it helped shape the way I see the world. Sciam Always educate me, joy, and inspire us to our vast and beautiful universe. I hope I do that for you too.

you Subscribe to Scientific Americanyou ensure that our coverage is centered around meaningful research and discovery. Having resources to report decisions that threaten labs across the United States. And we support both budding and working scientists when the value of science itself is not recognized too often.

In return, you get essential news, A fascinating podcast, great infographics, Miss newsletter, must-see videos, must-see videos, Challenging games and the world's best writing and reporting on science. You can do it too Give someone a subscription.

There was no more important time for us to stand up and show why science is important. I hope that you will support us on that mission.